A Hybrid Approach to Recommend Long Tail Items Diogo Vinícius de Sousa Silva Frederico Araújo Durão Federal University of Bahia Federal University of Bahia Salvador, Bahia, Brazil Salvador, Bahia, Brazil [email protected] [email protected] ABSTRACT would be in long tail) are responsible only for the minority of sales. Techniques in recommendation systems generally focuses on rec- According to Paretto’s rule [11], 80% of the consequences come ommending the most important items for a user. The purpose of this from 20% of the causes, that is, 80% of sales would be concentrated work is to generate recommendations focusing on long tail items, in only 20% of the products from the stock. Long tail items are those and then to conduct the user to less popular items. However, such less popular items, while the other items, i.e. the most popular, will items are of great relevance to the user. Two techniques from the be called “short tail” items. literature were applied in this study in a hybrid way. The first tech- The long tail phenomenon can also be seen as a way to increase nique is through markov chains to calculate node similarity of a user company profits. Generally high popular products are quite mar- item graph. The second technique applies clustering, where items keted by several other companies and of course the competition are separated into distinct clusters: popular items (short tail) and for sales is great. Since there is a great demand for these products, non-popular items (long tail). Using the Movielens 100k database, the price tends to be the lowest due to competition. Therefore, the we conducted an experiment to calculate the accuracy, diversity, products profit rate becomes quite low. Considering items with low and popularity of the recommended items. With our hybrid ap- demand, it is possible to predict a higher profit margin for interested proach we were able to improve the recall by up to 27.97 % when users. The users will be more motivated to pay a higher price due compared to the markov chain-based algorithm, which indicates to the low product availability. Another effect of exploring long tail greater targeting to long tail products. At the same time the rec- products is the so-called “one-stop shopping convenience” effect. ommended items were more diversified and less popular, which A store that offers long tail products and also popular products indicates greater targeting to long tail products. delivers an additional convenience for their customers since you find everything you need in one place. KEYWORDS In this context, this work proposes a hybrid technique for improv- ing the recommendation of long tail products. Our approach uses a Recommender System; Long Tail; Graphs; Markov Chain; Clusteri- structure that represents the items and users in a bipartite graph. zation. Then, we determine the proximity through Markov chains using the Hitting Time algorithm. In addition, we also combine the use of this 1 INTRODUCTION algorithm with a technique of splitting and clustering in order to Most of the methods used by Recommendation Systems (RS) tend improve the accuracy of long tail items recommendations. Finally, to recommend the most popular items to users. Since the majority we indicate the long tail items for recommendations, thus turning of users are interested in an item, the likelihood is that a new user the algorithm more assertive and improving the recommendations will also be interested in that item [8]. By following this logic, it diversity. is natural that less popular items are less recommended and con- The remainder of the work is organized as follows: Section 2 sequently less consumed. Usually, the main companies focus their presents the related works. The Section 3 addresses the proposed sales on these products envisioning better logistic. If we imagine long tail recommendation. Section 4 details the experimental eval- a company with physical stock, it is easy to understand that it is uation of the proposed technique and discusses the results of the much cheaper to put the best-selling products on the more evident experiments. Finally, Section 5 traces some conclusions and future shelves. However, with the advent of virtual stores the cost of orga- work. nizing products on shelves is non-existent. In the context of virtual stores has raised the term “infinite-inventory” [1], where products 2 RELATED WORK in evidence can be selected according to the preference of each Yin et al. [12] has developed four variations of an algorithm for long online user. This virtual user will not necessarily have the same tail item recommendations. The basic algorithm of the proposal is preference as other users. the Hitting Time, where the users and items are represented in a The term “long tail” refers to the set of products not commonly disjunct, indirect and bipartite graph. From this graph an adjacency consumed by users [1]. Usually, these products make up the bulk of matrix is obtained, as shown in Figure 1. The edges of the graph are store stock with low demand. Only a small amount of products con- weighted and represent the relevance of a user’s connection to an tribute for the majority of sales. In contrast, most products (which item, that is, a user’s rating for that item. To calculate the proximity In: XVIII Workshop de Teses de Dissertações (WTD 2018), Salvador, Brasil. Anais do of unrecorded items, the author calculates the Hitting Time using XXIV Simpósio Brasileiro de Sistemas Multimídia e Web: Workshops e Pôsteres. Porto a type of Markov chains called Random Walk [2]. Using Random Alegre: Sociedade Brasileira de Computação, 2018. Walk, the probability of a user reaching an item not evaluated by © 2018 SBC – Sociedade Brasileira de Computação. ISBN 978-85-7669-435-9. him is calculated. The higher the probability, the lower the Hitting Time, and therefore the item should have higher priority in the WebMedia 2018: Workshops e Pôsteres, Salvador, Brasil D.V.S Silva and F.A. Durão recommendation. Transition matrices are derived from a probabil- ity matrix. The probability matrix is obtained from the adjacency matrix shown in Figure 1. Shang et al. [10] perform a study of a custom recommendation model using collaborative filtering and ternary relationships, through tripartite graphs, representing users, items, and tags. It proposes a new measure for user similarity based on user tags and preferences. The similarities are calculated using a diffusion-based process and finally compared with recommen- dations calculated based on similarity of the cosine. Johnson et al. [4] carry out an extension of the work of Yin et al. [12]. John- son combines the algorithms used by Yin, adapting to tripartite graphs, shown by Shang. Johnson combined the Yin approach with the study of Shang to generate recommendations by collaborative filtering. Park and Tuzhilin [7] suggest an approach based on splitting and clustering item set. The item set is divided between short tail and various parts of the long tail. Thus, recommendation of short tail items are made based on the individual scores of each item. In Figure 1: Representation of users and items through a bipar- the case of the long tail items recommendation, the clusters scores tite graph and its respective adjacency matrix [12]. are grouped in each part of the long tail. Park in [6] evolves the previous work by clustering the items based on popularity. The author presents a technique called adaptive clustering for recom- mendations of items according to their popularity. It is possible to walk. A random walk has a current state (a node in the graph) and define the size and quantities of the clusters in an adaptive way with each evolution in time the state is changed, that is, visiting according to the state of the dataset, focusing on items long tail. As other nodes of the graph. Through the hitting time this path is a result you get lower error rates and improved performance. driven by the weights of the edges and the random walk ends its The approach presented in this work uses Hitting Time combined walk when it arrives at the target node. The algorithm is based on with a technique for item clustering. Our proposal will be based a probability matrix that is computed from the adjacency matrix. ¹ ¹ º on the Park’s approach [7]. We will group the long tail items into Finally, the edge weights are given by Equation pi; j = P s t + 1 = a¹i; jº Ín jjs¹tº = iº = , where di = a¹i; jº. clusters according to the average ratings of each item and use di j=1 this score as weight. The weight represents an additional variable Taking into account the probability matrix, a transition matrix in the Hitting Time algorithm. The main objective is to increase is calculated for each chain state until the destination state is reach. the emphasis on the recommendation of items located in long tail Next, the hitting time value may be obtained through of the H¹qjjº = Ín ¹ ; º improving the accuracy. 1 πj j=1 a i j = , where πi = Ín . pj;q pq; j πq i; j=1 a¹i; jº The Hitting Time algorithm is based on the time-reversibility 3 PROPOSED SOLUTION property to guide recommendations with long tail items. This prop- In this section we will explain our hybrid approach composed of erty indicates that the paths are not symmetric. The probability Hitting Time algorithm and a clustering technique will turn the that a node A reaches a node B is different from the probability recommendations more focused on long tail items. that node B reaches node A [2].